菜单
Forecast/predict future values of a time series, choose reasonable baselines, and compare forecasting methods on arbitrary series loaded from ts-agents data.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | time-series-forecasting |
| description | Forecast/predict future values of a time series, choose reasonable baselines, and compare forecasting methods on arbitrary series loaded from ts-agents data. |
| compatibility | Best with the ts-agents repo + CLI (`ts-agents`). |
| metadata | {"domain":"time-series","tasks":["forecasting","prediction","model-selection"],"ts_agents":{"tool_category":"forecasting","prefers_with_data_tools":true,"preferred_workflow":"forecast-series","preferred_tools":["detect_periodicity_with_data","forecast_seasonal_naive_with_data"],"artifact_checklist":["forecast_comparison.json","forecast.json","forecast.csv","report.md"]},"claude_code":{"allowed-tools":["Bash","Read","Write","Edit"],"disable-model-invocation":false}} |
Given a univariate time series, produce future predictions with a workflow that is:
If the user wants a reproducible benchmark/report workflow on the vendored M4
Monthly mini-panel, switch to SKILL-pro.md. This base skill
is for arbitrary run_id + variable series, not the fixed workflow contract.
run_id + variable, or raw array in Python)Useful extras:
If the season length is unknown and seasonality matters, estimate it first.
forecasting extra.forecasting extra.forecasting extra.forecasting extra.uv run ts-agents tool run detect_periodicity_with_data --run <RUN_ID> --var <VARIABLE> --param n_top=3
uv run ts-agents workflow show forecast-series --json
Interpret available_methods / unavailable_methods before you choose models.
In a default/base install, expect seasonal_naive to be available and
theta, ets, arima to remain unavailable until the optional
forecasting extra is installed.
uv run ts-agents workflow run forecast-series \
--run-id <RUN_ID> --variable <VARIABLE> \
--horizon 50 \
--methods seasonal_naive \
--season-length <PERIOD>
If you only need a quick point forecast instead of the workflow artifact set:
uv run ts-agents tool run forecast_seasonal_naive_with_data \
--run <RUN_ID> --var <VARIABLE> \
--param horizon=50 --param season_length=<PERIOD>
Notes:
workflow show forecast-series --json lists theta, ets, or arima
in available_methods, rerun the same workflow with those method names
added to --methods.unavailable_methods, stay with
seasonal_naive or install ts-agents[forecasting].Treat ensembles as an advanced path, not a base-profile default. Only use them after individual methods are available and have been compared explicitly.
The core forecasting functions support level=[80, 95], while the
*_with_data wrappers are mainly geared toward point forecasts and plots.
If you need interval-capable statistical models such as ETS, Theta, or ARIMA,
install ts-agents[forecasting] and use Python directly:
from ts_agents.data_access import get_series
from ts_agents.core.forecasting import forecast_ets
y = get_series("Re200Rm200", "bx001_real")
res = forecast_ets(y, horizon=50, level=[80, 95], season_length=12)
print(res.lower_bound[:5])
print(res.upper_bound[:5])
Return:
For stakeholder-facing outputs, produce a short Markdown or Quarto report with:
Convert ts-agents workflow outputs, run manifests, plots, JSON, CSV files, and short Markdown reports into stakeholder-facing Markdown, Quarto, HTML, or PDF reports. Use when the user asks for a report, executive summary, research appendix, or reusable analysis deliverable from existing ts-agents artifacts.
End-to-end workflow for labeled-stream activity recognition: prepare or download data, run window-size selection, evaluate a windowed classifier, and produce plots + a short report. Use when you need a reproducible CLI workflow artifact or evaluation bundle.
Supervised time series classification: choose and run classifiers (KNN/DTW, ROCKET variants, HIVE-COTE), compare models, and report accuracy. Use when the user asks to classify/categorize time series, build a classifier, or compare time series classification algorithms.
Decompose a time series into trend/seasonal/residual components (STL, MSTL, Holt-Winters). Use when the user asks about trend, seasonality, detrending, or wants residuals for anomaly detection/forecasting.
Quick EDA and diagnostics for a time series: descriptive stats, autocorrelation, and periodicity. Use when the user asks "what does this series look like?", "is there seasonality?", "what's the period?", or before choosing decomposition/forecasting parameters.